https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77834
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Created attachment 39960 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=39960&action=edit gcc7-pr77834-2.patch Per ml and IRC discussions, this patch attempts not to consider call mem operand as memory read from the first byte of a function in DSE. Unfortunately, it breaks pr67960.c testcase at -m32 -Os, we optimize away then the argument store; while it is store to 8(%ebp), for whatever reason cselib thinks it is %argp based rather than %ebp and as check is a const sibcall, doesn't consider the argument memory slot being needed. I'd say it is a RA bug, in that there is: (insn 16 3 6 2 (set (reg/v/f:SI 0 ax [orig:89 pi ] [89]) (mem/f/c:SI (plus:SI (reg/f:SI 6 bp) (const_int 8 [0x8])) [2 pi+0 S4 A32])) "pr67690.c":21 82 {*movsi_internal} (expr_list:REG_EQUIV (mem/f/c:SI (reg/f:SI 16 argp) [2 pi+0 S4 A32]) (nil))) (insn 6 16 7 2 (set (reg:SI 0 ax [orig:90 *pi_3(D) ] [90]) (mem:SI (reg/v/f:SI 0 ax [orig:89 pi ] [89]) [1 *pi_3(D)+0 S4 A32])) "pr67690.c":21 82 {*movsi_internal} (expr_list:REG_EQUIV (mem:SI (reg/f:SI 16 argp) [0 S4 A32]) (nil))) after reload - note in the first insn LRA properly reloaded the argp in (mem (argp)) in the SET_SRC, but not in REG_EQUIV. Vlad, do you agree that is a LRA bug? But not sure if fixing the REG_EQUIV would help, I've tried to update them from the debugger and df_insn_rescan at the beginning of dse2, but dse2 still removed the insns.
